add 4 medium models

[mike-ferguson]

Summary

• Adds four new Brain-Score language model plugins:
  • Qwen2.5-3B (Qwen/Qwen2.5-3B) — 36 layers, hidden size 2048, layer mapping model.layers.35
  • Falcon-7B (tiiuae/falcon-7b) — 32 layers, hidden size 4544, layer mapping transformer.h.31
  • OPT-6.7B (facebook/opt-6.7b) — 32 layers, hidden size 4096, layer mapping model.decoder.layers.31
  • Pythia-12B (EleutherAI/pythia-12b) — 36 layers, hidden size 5120, layer mapping gpt_neox.layers.35
• All models are registered via HuggingfaceSubject with the last transformer layer mapped to the language system, following the existing plugin pattern (e.g. Gemma).
• Each plugin includes tests for model loading, neural representation shape, and next-word prediction.

Test plan

• pytest brainscore_language/models/qwen2_5_3b/test.py -m memory_intense — all tests pass, neuroid dim = 2048
• pytest brainscore_language/models/falcon7b/test.py -m memory_intense — all tests pass, neuroid dim = 4544
• pytest brainscore_language/models/opt6_7b/test.py -m memory_intense — all tests pass, neuroid dim = 4096
• pytest brainscore_language/models/pythia12b/test.py -m memory_intense — all tests pass, neuroid dim = 5120